aidebate

aidebate

Multi-agent AI debate orchestrator. Spins up Claude Code, Gemini CLI, Codex CLI (and anything else with an interactive REPL) inside tmux panes, runs a structured debate between them, and serves a live web UI plus an archive browser.

Why

You have subscriptions to several AI coding assistants and occasionally want to pit them against each other — or let them cross-examine each other — on a concrete question ("Rust vs Go for this CLI", "Which of these 5 libraries should we pick?", "Is this migration worth it?"). Doing this manually is tedious; aidebate wires it up end to end:

• Each agent is a long-lived REPL in its own tmux pane (so you can watch them think live, or just see the final verdict).
• The orchestrator coordinates turns via flag files — agents never step on each other's keystrokes.
• A moderator agent runs the same way, reads everyone's answers at the end, and writes a verdict.
• Everything is persisted to disk as rendered Markdown, so every debate leaves a permanent, browsable transcript.

Install

You need tmux, Python 3.10+, and at least one authenticated AI CLI:

• Claude Code — run claude once and log in.
• Gemini CLI — run gemini once and sign in.
• Codex CLI — run codex login.

Then:

# One-shot (npx-style) with uv:
uvx aidebate serve

# Or persistent install:
pipx install aidebate
debate serve

Or develop from source:

git clone https://github.com/orlenko/aidebate && cd aidebate
python -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/debate --help

Quickstart

Web UI (recommended)

debate serve     # http://127.0.0.1:8765

Fill in the form — topic, moderator engine, and one row per debater (role + engine + stance). Click Start. The tiled tmux session spins up in the background and every pane streams into the browser. When it finishes, the verdict renders as proper Markdown and every run joins the "Prior debates" list for future reference.

Command line

debate run \
  --topic "Should we migrate from Framework A to Framework B?" \
  --side pro@claude:"Argue for migrating." \
  --side con@gemini:"Argue for staying." \
  --side auditor@codex:"Neutral cost analysis — no advocacy." \
  --moderator claude \
  --watch

--watch pops open an iTerm2/Terminal window already attached to the debate's tmux session so you can follow along. Without it, use the printed tmux attach -t debate-<timestamp> command.

Other subcommands

What gets saved

Each run writes to ~/.aidebate/sessions/<id>/. Override with AIDEBATE_HOME=<dir> or --sessions-dir <dir> on any subcommand. (We deliberately avoid macOS's ~/Library/Application Support/ path — the space breaks too many of the ad-hoc shell commands that AI agents emit mid-debate.)

~/.aidebate/sessions/2026-04-14-120000/
├── session.json                    # manifest: topic, sides, moderator, status, timings
├── chat.jsonl                      # group chat log (one JSON object per line)
├── verdict.md                      # moderator's final ruling
├── agents/
│   ├── pro/                        # each agent's cwd
│   ├── con/
│   └── moderator/
├── phase-1-opening/
│   ├── pro.prompt.md
│   ├── pro.answer.md
│   ├── pro.done
│   └── …
├── phase-2-rebuttal/
└── phase-3-verdict/

How it works

Agents coordinate via flag files, not terminal scraping:

1. Orchestrator writes phase-N/<role>.prompt.md.
2. It tells the agent (via tmux send-keys) to read that file and follow it.
3. The prompt instructs the agent to write its response to <role>.answer.md and then touch '<role>.done'.
4. Orchestrator polls for the .done flag; when every agent's .done lands, the phase is complete.

The debate runs in three phases:

Phase 1: Opening    — each debater writes their position (parallel)
Phase 2: Rebuttal   — each debater attacks every opponent's opening (parallel)
Phase 3: Verdict    — moderator reads all answers + chat, renders ruling

startup_keys in the adapter config auto-dismiss "trust this folder?" dialogs for each CLI. permission_prompts handle anything else interactive.

Adapters

Adding a new AI CLI is a YAML file under src/aidebate/adapters/:

# src/aidebate/adapters/mynewcli.yaml
name: mynewcli
cmd: "mynewcli --yolo --workspace {session_root}"
submit_key: "Enter"
submit_delay: 0.8
permission_prompts:
  - { match: "Trust this", respond: "" }  # empty respond → just press Enter
  - { match: "\\[y/N\\]",  respond: "y" }
answer_instruction: |
  When you have finished this task, write your full response to:

      {answer_path}

  Then run:

      touch '{done_path}'
startup_keys:
  - { delay: 3.0, key: "Enter" }   # dismiss startup dialogs

The adapter's cmd can reference {session_root} and {agent_cwd}, substituted at spawn time.

Development

.venv/bin/pytest                # run the test suite
.venv/bin/ruff check src tests  # lint
scripts/hostile-eyes            # adversarial self-review + mark current HEAD
scripts/pre-pr-check            # fail fast if hostile review is missing
scripts/restart-server          # bounce the local web server on :8765

The tests don't require tmux or any AI CLI — they cover adapter loading, side parsing, prompt building, and the FastAPI endpoints that read session state from disk. For end-to-end runs you need the real thing.

License

Apache-2.0. See LICENSE.